OpenClaw

Jue status: Read, Write, and Confirm are Implemented (packages/ai-jue-adapter-openclaw/). workspace and compatible-bundle (RFC-0002) both ship. The latter delegates to Claude/Codex plugin writers and follows OpenClaw's official Bundle discovery path. The OpenClaw native runtime surface remains an Agent-specific Artifact surface in the official capability table; Jue's Canonical mapping focuses on workspace and compatible-bundle. Workspace capabilities honestly declare rules/commands/agents/mcp: "degraded".

Official sources: Plugin bundles, Plugins, Building plugins, Plugin Manifest

1. Official surfaces (two layers)

1.1 Workspace (project tree)

AGENTS.md, skills/<name>/SKILL.md, hooks/<name>/HOOK.md+handler.js.
OpenClaw's global Agent surface manages commands/agents runtime entry points; the project tree contains AGENTS.md, skills, and hooks, while MCP lives in the global openclaw.json.

1.2 Installable plugins (current docs)

Format	Marker	Role
Compatible bundle	.claude-plugin/ / .codex-plugin/ / .cursor-plugin/ (or default Claude layout)	Content packs; mapped skills/hooks/MCP; narrow trust boundary
Native plugin	openclaw.plugin.json + package.json#openclaw.extensions	In-process tools/channels/providers

openclaw plugins install ./my-bundle
openclaw plugins list    # bundles show Format: bundle + Bundle format
openclaw plugins inspect <id>

Jue's compatible-bundle detection reads the Bundle markers above; the native runtime surface is handled independently by OpenClaw's official loader.

Bundle mapping highlights:

• skills: all formats
• Claude/Cursor commands/ → skill roots
• hooks: only OpenClaw-style HOOK.md+handler execute (mainly Codex-compatible packs); Claude hooks/hooks.json is detect-only
• agents: often detect-only for Claude/Cursor
• MCP: merged into embedded settings

2. Intended Jue mapping

Canonical / Adapter duty	OpenClaw
context.global	Workspace: AGENTS.md (managed block)
skills	Workspace: skills/<name>/; Bundle: Claude/Codex plugin layout; both preserve references/, scripts/, and assets/ support files
hooks	Workspace: HOOK.md+handler.js; Bundle: prefer Codex base to execute
commands / agents / mcp.servers	Workspace: honest degraded; Bundle: per official mapping / detect-only
Artifact	workspace | compatible-bundle (RFC-0002)
Confirm	Workspace: openclaw --profile … config validate; Bundle: plugins install + inspect (Format: bundle)

3. Conversion boundary

• Workspace MCP stays outside the project-level Artifact, protecting the user's global openclaw.json.
• compatible-bundle reuses official Claude/Codex layouts, or an explicit bundleFormat: "cursor" Cursor artifactKind: "plugin" layout. auto never selects Cursor.
• Canonical packs keep the Bundle's narrow trust boundary; the OpenClaw native runtime surface stays described by the Agent's official capability table.
• When hooks run under OpenClaw, choose the Codex bundle base; without runnable hooks, the Claude base is the default.

4. Current gaps

Level	Status	Gap
Read / Write / Confirm (workspace)	Implemented	packages/ai-jue-adapter-openclaw/
Artifact compatible-bundle	Implemented	Delegates to Claude/Codex artifactKind: "plugin"; install confirm needs the local openclaw CLI (skipped when it is not installed)
Native plugin Artifact	Unsupported	Explicitly out of scope (OpenClaw's official loader handles it); official runtime surface documented separately from Canonical bundles (not a gap)